Upload
others
View
1
Download
0
Embed Size (px)
Citation preview
Perceptions of Equity and the Distribution of Income �
Julio J. Rotemberg
Revised October 20, 2000
Abstract
This paper builds a simple model in which there is a connection between employ-
ees' perception of the \fairness" of employers and the actual distribution of income.
Employers base their wages on their assessments of the productivity of individual em-
ployees. I show that the equilibrium distribution of income depends on the beliefs of
employees concerning the accuracy of these assessments. This distribution tends to be
more dispersed the more accurate employees believe these evaluations to be. This �ts
with the observation that, in a sample of seven countries, there is a negative correlation
between actual inequality and the extent to which inequality is perceived to be exces-
sive. The changes in beliefs that increase inequality in the model can be expected to
lead voters to favor candidates who oppose redistribution. The model can thus account
for the disproportionate increase in inequality in countries, such as the United States
and the United Kingdom, where the popular political discourse has shifted against
redistribution. The model is also consistent with some recent changes in job tenure
in the United States and France, where the former experienced a much larger increase
in inequality. Among highly educated workers, the tenure of workers whose wages are
relatively high has been falling relative to that of lower-paid workers in the United
States, while the opposite is true in France.
�Graduate School of Business Administration, Harvard University I am very grateful to three anonymous
referees as well as Mauro Guillen, Frank Levy, Mancur Olson, and Michael Woodford and the participants
at seminars at Columbia, Harvard, Johns Hopkins, the NBER and the University of Maryland for helpful
suggestions. I also wish to thank Hank Farber and Dan Feenberg for help with the CPS, Francis Kramarz for
providing me with data that he obtained by processing the Enqu�ete Emploi and the NSF and the Harvard
Business School Division of Research for research support.
People have opinions about the degree to which di�erences in pay re ect di�erences in
productivity or are, instead, the result of favoritism, personal connections and the whimsy
of people with in uence. In this paper, I show that opinions of this sort can matter a great
deal. In particular, I demonstrate that opinions about the extent to which people believe
that di�erences in pay re ect di�erences in individual productivity can a�ect the actual
distribution of income even when they have no e�ect on individual marginal products.1
My focus on the e�ects of these beliefs raises the question of where these beliefs come
from. In particular, there is the issue of whether one shouldn't simply assume that people
know the truth so that there is no di�erence between beliefs about the connection between
income and productivity and the actual connection between these variables. In this paper,
I allow the actual and the perceived connections to di�er and I do this for two reasons.
First, the truth about the connection between pay and productivity is diÆcult to know.
If one leaves aside the possibility of experimenting by varying one's own productivity, it is
very hard to learn about this connection unless one actually observes the output of di�erent
people.2 But, in modern industrial �rms where workers are extremely interdependent and
where the aggregate value produced by the �rm is subject to large random disturbances,
measuring the output of individuals is very hard. There are, of course, indirect measurements
of the relationship between productivity and pay. Under the assumptions of the stripped
down model I present, for instance, this relationship can be inferred from the statistical
properties of individual wages. However, the extent to which statistics of this sort shed
light on the issue depends on auxiliary assumptions, the validity of which is itself diÆcult to
establish empirically.
A second, related, reason to consider models where the perceived connection between
pay and productivity di�ers from the actual one is that people do not appear to agree on
1The paper is closely related to Piketty (1995) where changes in the perceived connection between e�ort
and income change the equilibrium level of e�ort and thus change actual income. The di�erence is that in
Piketty (1995), these changed beliefs a�ect actual income only by �rst a�ecting the potential productivity
of individuals.2This too is closely related to Piketty's (1995) argument that it is diÆcult to learn the connection between
e�ort and pay.
1
this connection. A particularly telling set of examples of this disagreement can be found in
Hochschild (1981), who reports on twenty eight in-depth interviews about the distribution
of resources. Summarizing she says (p. 140), \The poor . . . often argue that if productivity
were truly rewarded, this would create more equal incomes (italics in original)." And goes on
to say (p. 141) \On the other hand, wealthy respondents . . . often argue that if productivity
were truly rewarded, this would create less equal incomes." 3
The diÆculties in determining individual output do not stop �rms from basing their pay
on estimates of this output. Indeed, in the case of a frictionless competitive labor market, the
wage of an individual is equal to his employer's estimate of his marginal product. Opinions
about whether people are or are not paid their marginal product do not alter this so that
they have no e�ect on the actual distribution of income.
In the model I present, there are two departures from the assumptions of the standard
competitive labor market which ensure that these opinions matter. First, I suppose that a
worker's departure from a �rm depends on his outside opportunities and that the �rm has
imperfect information about these opportunities. The result is that employers have some
monopsony power and that wages depend both on the �rm's perception of their employees'
marginal products, and on the �rm's assessment of the likelihood that these employees will
leave. Second, there are search costs so that workers have to form an opinion about what
they will be paid on the outside before they leave a �rm.
This combination of assumptions yields a simple mechanism by which opinions a�ect the
distribution of income. If workers regard the evaluations of their employers as inaccurate,
they are tempted to quit if they receive a low evaluation while they are inclined to stay if
they receive a favorable one. The reason is that, if �rms are inaccurate, it is good to work
for a �rm that has an inaccurately high estimate of one's own ability. Employees who receive
high ratings from their current employers conclude that they have found a good match while
3These di�erences are consistent with those found in larger surveys. McClosky and Zaller (1984) report
that 84% of respondents whose family income is above $35000 view the capitalist system as fair and eÆcient
while only 51% of respondents with lower income do so. This type of statement may simply re ect di�erent
tastes for redistribution but it also could re ect di�erences in opinions about the determinants of individual
incomes.
2
those who receive a low rating expect alternate employers to pay higher wages.
Thus, if workers change their minds and come to see �rms as being more accurate, they
change their quitting behavior. Workers who receive a low rating from their �rm become
more predisposed towards staying (because they now expect a low rating from other �rms
as well). A rational monopsonistic �rm would then tend to cut their wages. Similarly, this
change in perception leads workers that have gotten a high rating to become more prone to
quit because their fear of getting low wages from alternate employers is reduced. This, in
turn, tends to raise their wages. Since the wages of highly rated employees rise while those
of employees with low ratings fall, inequality increases.
The paper proceeds as follows. Section 1 presents the model and studies how the wage
distribution changes in response to changed perceptions by workers of the accuracy of �rm's
evaluations. Section 2 studies the change in aggregate output when workers come to perceive
�rms as more accurate. One diÆculty with evaluating aggregate outcomes is that one needs
an \objective" measure of �rms' ability to evaluate workers and, for the reasons I mentioned
it is hard to know how one would go about obtaining one. For the purpose at hand, I
pretend that �rms have an accurate assessment of their own ability to judge individuals'
productivity. This leads me to use the �rms' estimates of their accuracy in computing the
model's predictions regarding aggregate output.
The main conclusion of this section is there are plausible conditions under which ag-
gregate output declines when workers deem �rms to be more accurate, though there also
exist conditions under which output rises. One reason for output to decline is that increased
con�dence in the ability of �rms to evaluate workers leads more high-wage workers to quit
and there is often a larger deadweight loss associated with these quits than there is with the
quits of low-wage workers.
Section 3 discusses the empirical relevance of the model. Beliefs about the fairness of the
distribution of income di�er both cross-sectionally and over time. In particular, residents
of di�erent countries harbor di�er beliefs about this fairness. I thus study the connection
between the income distribution in individual nations and the beliefs held in those nations.
3
I show that, across a sample of seven countries, actual inequality is negatively correlated
with the perception that income inequality is excessive. This is in broad accordance with
the predictions of my model.
It also may be possible for beliefs about the fairness of the income distribution to change
fairly rapidly within a country. This could occur as a result of the spreading of incontrovert-
ible new evidence (as in Bolton and Harris 1999) but might also happen in cases where the
\new" evidence is relatively tenuous. The perception that inequality is excessive is presum-
ably associated with voting for candidates who favor redistribution. Thus, changes in the
dominant political discourse can to some extent be read as changes in the degree to which
individuals see unequal outcomes as fair. The model would then be able to account for the
fact that income inequality has increased disproportionately in countries such as the U.S.
and the U.K. where the political support for redistribution appears to have waned.
Since the implications of the model for the distribution of income are closely tied to
its implications for quits, it seems worthwhile to compare the changes in job tenure across
countries who have experienced di�erent changes in their income distributions. I thus study
changes in job tenure in France and the United States. According to the model, job tenure
among those with relatively high incomes should have fallen in the United States relative to
the tenure of lower-wage individuals (as it is the willingness to hop from job to job among
high-wage individuals that has raised their wages). By contrast, in France, job tenure among
individuals with high incomes should not have fallen relative to that of individuals with
low incomes. The actual changes in the tenure of individuals with relatively high levels
of education are consistent with these implications, though this is not true for individuals
whose educational attainment is lower. Section 5 concludes.
4
1 The Model
1.1 The Basic Structure
Employees are of two types. I suppose that those for whom a parameter r is equal to one
have higher productivity, at least in certain occupations, than those for whom the parameter
r is equal to zero. While �rms do not observe r directly, they do observe signals that are
correlated with r. The focus of my study is the connection among wages, employee turnover
and employee beliefs about the accuracy of �rms' signals about r.
There are two periods because this is the minimal number of periods required to study
turnover and my model is one where beliefs a�ect wages only through their e�ect on the
tendency to quit. The sequence of potential employers for each employee is depicted in
Figure 1. In each period, employees choose whether to work at a �rm of type S where their
output equals qS + vT with v > 0, or whether to work at a �rm of type N where employee
output is independent of r. Individuals start the �rst period attached to particular �rms of
type S in the sense that, for each employee, there exists a �rm which observes a signal s1
which is correlated with the employee's productivity. This �rm makes a wage o�er to the
employee at the beginning of period 1. Simultaneously, the employee receives an o�er from
a type N �rm. The employee then chooses which �rm he wants to work for in period 1.
If he works for the �rm of type N in period 1, there is a second type S �rm which
observes a new signal s2 about the employee's productivity at the beginning of period 2.
This employee then receives two wage o�ers in period 2, one from this new type S �rm and
one from a new type N �rm. If, by contrast, the employee stays with his initial type S �rm
in period 1, no other type S �rm o�ers him a job in period 2. Such an employee must thus
decide in period 2 whether to accept the new wage o�er that he receives from his initial
type S �rm or whether to accept an o�er made by a new type N �rm. This captures two
aspects of quitting a job. The �rst is that departure from a job often expands an employee's
range of opportunities. The second is that the opportunities that will become available are
somewhat uncertain at the time the individual quits.
5
The o�ers from �rms of type N play a dual role in my analysis. On the one hand,
they dampen the monopsony power of the type S �rms and this is indeed their only role in
period 2. In addition, they serve as transitional jobs for employees who wish to move from
one type S employer to another. To keep the analysis simple I suppose that �rms of type
N pay employees the value of their marginal product qN while also conferring on to them
a nonpecuniary utility which is equivalent to an additional compensation of n.4 Abusing
the language somewhat, I thus treat him as receiving a total compensation of qN + n. For
any given employee, the value of n is drawn independently at each point in time from a
distribution whose cumulative density function is F (n) with 0 � n � �n.5
There is another interpretation for n which yields similar results for the behavior of wages
and quitting decisions. Instead of being thought of as a nonpecuniary bene�t of working at
a �rm of type N , n can be thought of as a nonpecuniary cost of working at a �rm of type
S. One can imagine that at the beginning of each period, the employer of type S assigns
a particular task to the employee and that this gives him a disutility of n. The analysis of
equilibrium wages in the second period is then unchanged as long as the level of this disutility
is unknown to the employer of type S. First period wages are not the same, but respond in
a similar way to changes in employee beliefs.
I analyze the model backwards starting with the wages paid in the second period at �rms
of type S. I then study expected compensation in the second period as viewed from period
one. This di�ers from expected wages at S because individuals have the option of working in
�rms of type N . Finally, the last subsection describes equilibrium wages in the �rst period.
4The simplest rationalization for this is there are actually two alternative �rms in which the individual
has a productivity qN and that the employee gets the same nonpecuniary utility in both. If each of these
�rms gets to make one wage o�er to the employee then the logic of Bertrand competition leads them to o�er
him qN .5Using zero as the lower bound for the realizations of n does not reduce the generality of the analysis
since one can think of the lowest value for n as being incorporated in qN .
6
1.2 Second Period Wages
Employers in type S �rms know that employees will leave and go to type N �rms if they are
o�ered a wage w that is lower than the employee's realization of qN + n (or, equivalently,
if qN exceeds the realization of w � n). The expected bene�t from keeping an employee to
which the �rm pays w is (qS + vEr � w), where Er represents the �rm's expectation of the
employee's r. Thus, these employers set their wage o�ers w to maximize
F (w � qN )[qS + vEr � w] (1)
where E takes expectations conditional on the information available to the �rm.
The �rst order condition for this problem is
f(w� � qN )[qS + vEr � w�]� F (w� � qN ) = 0: (2)
This represents the optimum wage as long as the boundary conditions
[f(0)[qS � qN + vEr] > 0 f(qN + �n)[qS � qN � �n + vEr] < 1 (3)
and second order condition
f0[qS + vEr � w
�]� 2f < 0 (4)
are satis�ed. I assume that f(0) is positive and that, for all workers,
qS � qN + vEr > 0 (5)
which ensures that the �rst boundary condition holds so that [qS + vEr � w�] is positive in
equilibrium and the �rm earns positive rents on employees who stay. If the rents [qS � qN +
vEr� �n] are suÆciently large, the second boundary condition in (3) is violated and the �rm
simply pays qN + �n which ensures that the employee stays with probability one.
Turning back to the interior case, di�erentiating (2) with respect to Er, qN and w� one
obtains
dw� =
f
2f � f 0[qS + vEr � w�]vdEr +
f � f0[qS + vEr � w
�]
2f � f 0[qS + vEr � w�]dqN : (6)
7
The second order condition (4) implies that the denominator in these expressions is
positive so that an increase in the employee's expected productivity r raises the wage o�er.
To ensure that increases in the employee's outside wage, qN , also raise the �rm's wage o�er,
it is necessary and suÆcient that
f0[qS + vEr � w
�]� f < 0 (7)
which is consistent with (4) but puts an even tighter bound on the extent to which f0 is
allowed to be increasing. This bound is needed because an increase in qN lowers the value
of n at which individuals depart for a given wage. If f 0 is increasing in n, this implies that
an increase in qN lowers the extent to which increases in w promote employee retention,
and this tends to reduce the optimal wage. Given (2), (7) is satis�ed as long as f 0
F� f2
F 2 is
negative. This, in turn, is satis�ed if and only if the hazard f=F (which is the hazard of
increased departures due to reductions in n) decreases as n rises. I assume this monotone
hazard condition from now on.
To compute the optimal wage, one must know the value of Er. For simplicity, I assume
that s1 and s2 are the only sources of information about r used by �rms. In particular,
�rms whose employees join them in period 2 ignore the work histories of these employees in
computing Er. This is so even though such �rms ought to recognize that employees who
left their period 1 employers have, on average, drawn di�erent values of s1 than those that
didn't. The simplest rationalization for this is to suppose that period 2 employers do not
believe that the signals observed by these earlier employers are informative.6 In an earlier
version of this paper I imagined instead that they computed Er under the assumption that
they regarded these signals as informative. While this complicates the analysis considerably,
it neither changes the results nor contributes much additional insight.
Along similar lines, I suppose that the work carried out by employees during period one
does not lead their own employers to acquire additional information about their productiv-
6It may also be possible to rationalize such an assumption by embedding this model into one where there
are overlapping generations of workers and �rms do not know the age of their employees. In such an extended
model, however, the quality of the pool of unattached workers is likely to depend on �rms' wage policies as
in Acemoglu and Pischke (1996) and I ignore this in my analysis.
8
ity. Again, the main reason for this assumption is the extra simplicity it brings, though
the assumption can be justi�ed by supposing that the �rm's only additional information is
an indicator of total output which contains negligible information about individual contri-
butions.7 Alternatively, one can imagine that the value of the employee's output accrues
over time and that the e�ect on �rm output of employees of di�erent productivities is not
detected until much after the employees have left the �rm.
I now turn to the characteristics of the signals s1 and s2. I let the unconditional proba-
bility that s1 (or s2) is equal to one as well as the unconditional probability that r is equal
to one be equal to �. Thus � denotes both the fraction of high productivity individuals and
the fractions of individuals who get high signals.
Consider now the beliefs concerning the likelihood that s1 (or s2) equals a particular
value as a function of the value taken by r. It is important to stress that these are subjective
beliefs concerning the accuracy of a signal. Thus these beliefs can di�er across agents. I
therefore denote by � and �w respectively the subjective conditional probabilities held by
�rms and workers that s1 (or s2) equals one conditional on r being equal to one. To ensure
that the signals are regarded as informative about r, I assume that both � and �w exceed �.
I let rz denote the subjective probability of �rms that r is equal to 1 given that s1 (or
s2) is equal to z. This also equals their conditional expectation of r. These conditional
probabilities are
r1 � P (r = 1js1 = 1) =
P (s1 = 1jr = 1)P (r = 1)
P (s1 = 1)= � (8)
r0 � P (r = 1js1 = 0) =
P (s1 = 0jr = 1)P (r = 1)
P (s1 = 0)=
(1� �)�
1� �: (9)
Letting wz2 denote the wage paid in the second period by type S employers to employees
whose s1 (or s2) is equal to z, this implies
Proposition 1: w12 > w
02
Proof: Because � exceeds �, (8) and (9) imply that Er (which equals rz) is larger when the
7Even if such an indicator contained good information about average productivity, this would not be very
useful if economies of scale obliged the �rm to hire many employees of independent ability.
9
signal is equal to 1. Given (6), this implies that the wage is higher as well.
At the same time:
Proposition 2: w12 � w
02 < v(r1 � r
0)
Proof Using (2) in (6),dw
�
dEr=
f
f + ff � f 0Ffg: (10)
Since the term in curly brackets is positive, this is smaller than one. This immediately
implies the result.
Thus, while employees with s1 = 1 receive higher wage o�ers than employees with s1 = 0,
the di�erence between an employee's expected productivity and his wage is also increasing
in the value of s1.
1.3 The Expectation of Second Period Compensation in Period 1
I suppose that each individual worker knows the signal s1 that is observed by his original
type S employer.8 Supposing n is the nonpecuniary compensation in a �rm of type N , an
employee with s1 = z who stays with his original employer can expect to earn wz2 if qN + n
turns out to be less than wz2. Otherwise he earns qN +n. Thus, such an employee can expect
to earn cz2 in the second period where
cz2 = F (wz
2 � qN)wz2 +
Z �n
wz2�qN
(qN + n)dF (n)
= qN + F (wz2 � qN)(w
z2 � qN) +
Z �n
wz2�qN
ndF (n) (11)
The derivative of cz2 with respect to wz2 is the probability of earning this wage, F (w
z2�qN ).
Because this is positive proposition 1 implies that the expected compensation of an employee
whose s1 equals to one exceeds that of an employee whose s1 is equal to zero. However,
8Alternatively, one can assume that he has to infer his s1, and thus his second period wage at both his
current and alternate type S �rm, from his �rst period wage. This gives the same equilibrium allocation
as in the case where the employee observes s1 directly as long as employees with di�erent values for s1 get
di�erent �rst period wages. What complicates the analysis of this case is that the �rm can now choose to
pay all employees the same �rst period wage so as to make it impossible for them to learn their realization
of s1. In numerical simulations I have carried out, I have not found any parameter combinations for which
the �rm would pro�t from this concealment of information. Thus, for the parameters I have studied the
equilibrium is the same whether employees observe s1 directly or not.
10
because this derivative is less than one, expected compensation does not rise as fast as this
wage o�er. The reason, of course, is that individuals who receive low wage o�ers are more
likely to leave their type S employer.9
Consider now the expected compensation of an employee who leaves his initial type S
employer. If this employee draws a signal s2 equal to z at his new type S employer in period
2, he can expect to receive a compensation of cz2. This means that his expected period 2
compensation is
czL = Pw(s2 = 0js1 = z)c02 + Pw(s2 = 1js1 = z)c12 (12)
where Pw represents employees' subjective probabilities. This compensation depends cru-
cially on the conditional distribution of s2 given s1. I calculate this under the additional
assumption that everyone agrees that the realization of s2 is independent of the realization
of s1 once one conditions on the true value of r. In other words
P (s2 = 1js1 = j; r = k) = P (s2 = 1jr = k) where k; j = 0; 1: (13)
The worker's subjective probability that a signal s1 equal to one will be followed by an
s2 equal to one is then
Pw(s2 = 1js1 = 1) = Pw(s2 = 1; r = 0js1 = 1) + Pw(s2 = 1; r = 1js1 = 1)
= Pw(s2 = 1js1 = 1; r = 0)Pw(r = 0js1 = 1) + Pw(s2 = 1js1 = 1; r = 1)Pw(r = 1js1 = 1)
=(1� �w)
2�+ �
2w(1� �)
1� �= �+
(�w � �)2
1� �: (14)
Similarly, the subjective probability that s2 will equal zero given that s1 equals zero is
Pw(s2 = 0js1 = 0) = Pw(s2 = 0; r = 0js1 = 0) + Pw(s2 = 0; r = 1js1 = 0)
= Pw(s2 = 0js1 = 0; r = 0)Pw(r = 0js1 = 0) + Pw(s2 = 0js1 = 0; r = 1)Pw(r = 1js1 = 0)
=h1�
�(1� �w)
1� �
i2+�(1� �
2)
1� �= 1� �+ �
��w � �
1� �
�2: (15)
9Consider instead the case where n is the disutility of working at the type S �rm. The employee then
earns wz2�n for n below wz2�qN and qN otherwise. Thus, his expected compensation is qN+F (w
z2�qN )(w
z2�
qN )�R wz
2�qN
0ndF (n). This di�ers from the value in (11) only by the average value of n,
R �n
0ndF (n) which is
a constant independent of wz2 . Because only the properties of the connection between expected compensation
and the second period wage matter for the comparative statics concerning how �rst period wages and turnover
respond to �w , these comparative statics carry over to this case as well.
11
Both the expression in (14) and that in (15) are increasing in �w (assuming the signal is
seen as informative so that �w exceeds �). The intuitive reason for this is that a higher �w
implies that both s1 and s2 are seen as more highly correlated with r. As a result, a higher
�w ensures that s1 and s2 are more highly correlated with each other.
Using (12), this implies that expected compensation for employees who leave with s1
equal to zero and one are, respectively
c0L = Pw(s2 = 0js1 = 0))c02 + (1� Pw(s2 = 0js1 = 0))c12
= c02 + �
h1�
��w � �
1� �
�2ihc12 � c
02
i(16)
c1L = (1� Pw(s2 = 1js1 = 1))c02 + Pw(s2 = 1js1 = 1))c12
= c02 +
h�+
(�w � �)2
1� �
ihc12 � c
02
i: (17)
This implies immediately
Proposition 3: An increase in �w raises c1L and lowers c0L.
Proof: Di�erentiating (16) and (17),
dc1L
d�w= �
1� �
�
dc0L
d�w= 2
�w � �
1� �[c12 � c
02] (18)
This is positive because �w > � and because the discussion below proposition 1 implies
that c12 exceeds c02.
Thus, the belief that �rms are more inaccurate (so that �w is low) leads workers who get
an s1 equal to one to expect lower wages on the outside while it leads workers who get an
s1 equal to zero to expect higher wages on the outside. The reason is that a low value of
�w leads workers to regard both the current realization of s1 and the future realization of s2
as likely to be mistaken. As �w falls, workers deem it more likely that a high realization of
s1 will be followed by a low realization of s2. Thus, an employee who gets a high s1 value
becomes less optimistic about his compensation on the outside. Similarly, an employee with
a low s1 becomes less pessimistic about his compensation by an alternate type S �rm.
12
Note that �w matters only because it a�ects workers' subjective distribution of the wages
they would be paid by their next employer if they quit their current job. Thus, �w would not
matter if workers had independent knowledge of the distribution of future wage o�ers condi-
tional on a worker's characteristics including their current wage. Obtaining this information
directly by moving frequently is costly, however. Aggregate statistics describing the wage
changes of job movers will generally not suÆce as these depend critically on the distribution
of n which a�ects the extent to which job changers started out with di�erent values of s1.
More generally, workers are unlikely to know the distribution of shocks, including shocks to
individual worker productivity from which I have abstracted, which prompt other workers
to move. This ignorance reduces the value of information about the wage distribution of
job changers. The result is that general opinions about the fairness of �rms, which I have
captured through the parameter �w, can play a role in workers' quitting decisions.
1.4 First Period Wages
I now turn to the analysis of the wages paid in period 1 by �rms of type S. I denote by wz1 the
wages paid by these �rms to the employees whose signal s1 equals z. Ignoring discounting, an
employee who is o�ered this wage and stays with his original type S employer can expect to
receive wz1+ c
z2 over the two periods. By leaving, the employee can expect to earn �qS+n+ c
zL
over the two periods. Thus employees whose s1 equals z stay if their realization of n is lower
than that the cuto� value nz1 such that
wz1 + c
z2 = qN + n
z1 + c
zL (19)
Thus employees stay with probability F (wz1 + c
z2� qN � c
zL) and the pro�ts of a �rm that
o�ers a wage equal to wz1 to an employee whose signal is z equal
F (wz1 + c
z2 � qN � c
zL)nqS + vr
z � wz1 + F (wz
2 � qN)[qS + vrz � w
z2]o: (20)
In this expression, the term in curly brackets takes into account that the presence of the
employee in the �rst period leads to a probability F (wz2 � qN ) that the employee also stays
13
in the second period. The �rm thus maximizes (20) with respect to wz1 while taking w
z2 as
given. Assuming an interior solution, the �rst order condition that characterizes wz1, is
f(wz1 + c
z2 � qN � c
zL)nqS + vr
z � wz1 + F (wz
2 � qN )[qS + vrz � w
z2]o
�F (wz1 + c
z2 � qN � c
zL) = 0: (21)
Di�erentiating (21) with respect to the wage and to czL, and using (21), gives
dwz1
dczL
=f2 � f
0F
2f 2 � f 0F(22)
where F , f and f 0 are evaluated at (wz1+c
z2��qS�c
zL). The second order condition ensures that
the denominator is positive. My assumption that f=F is uniformly decreasing in n implies
that the numerator is positive as well. An increase in expected outside compensation in the
second period leads then leads �rms of type S to raise their �rst period wage. This, together
with Proposition 3, implies immediately
Proposition 4: An increase in �w increases w11 and reduces w0
1.
Thus, an increase in �w increases the dispersion of the initial wages o�ered by �rms of
type S if w11 exceeds w
01. While the necessary conditions for this to occur are much weaker,
a suÆcient condition is given by the following proposition.
Proposition 5: If f 0 � 0, w11 > w
01
Proof: Let w01 be the optimal wage when s1 = 0. The optimal wage w1
1 exceeds this wage if
the left hand side of (21) is positive for z equal 1 when w11 is set equal to w
01. I demonstrate
that this is the case by holding constant w1 and cL, and showing that increases in vrz that
are accompanied by the corresponding increases in wz2 and c
z2 raise the left hand side of (21).
The monotonicity of f=F together with c1L > c
0L then ensure that the left hand side of (21)
evaluated at the point where w11 is equal to w
01 is indeed positive.
Let d denote the derivative of cz2 with respect to changes in vrz. This is given by the
expression in (10) times F (wz2 � qN) so that it is positive and less than one. Using the
envelope theorem for the optimality of wz2, the change in the left hand side of (21) as one
changes vrz while varying wz2 and c
z2 is
df0nqS + vr
z � w01 + F (w2 � qN )[qS + vr
z � wz2]o+ f(1 + F (wz
2 � qN)� d)
14
where f and f0 are evaluated at (w0
1 + c2 � qN � c0L). Since d < 1 this expression is positive
if f 0 is nonnegative, though this condition is clearly stronger than necessary.
The uniform distribution simultaneously satis�es the condition that f=F be monotoni-
cally decreasing in n and the condition that f 0 be nonnegative. Thus, the gap between w11
and w01 is increasing with �w for this distribution. This gap corresponds, broadly, to the
variability of wages inside �rms. Consider next the gap between the highest and lowest �rst
period wage paid in the model economy, which corresponds broadly to measures such as
the interdecile range. Whether qN represents the lowest, the middle, or the top wage, this
measure of inequality rises with �w as long as w01 falls and w
11 rises.
The model also has implications for turnover, and I turn to these next. Di�erentiating
(19)dn
z1
d�w= �
dczL
d�w
1�
dwz1
dczL
!=
�f 2
2f 2 � f 0F
dczL
d�w(23)
where the second equality follows from (22). This means that, in spite of the corresponding
increase in wz1, a rise in c
zL unambiguously lowers the cuto� n
z1 at which employees leave and
thereby raises the number of quits. Therefore, a reduction in �w increases the turnover of
employees whose s1 is equal to zero while it lowers the turnover of employees whose s1 is
equal to one.
I now provide some intuition for the result that increases in �w raise the relative turnover
of high s1 individuals together with their relative wages. Figure 2 helps one interpret these
results. This �gure shows the labor supply curve given, implicitly, by (19). This is a labor
supply curve because a higher wage leads workers to have a higher cuto� nz1 so that the
fraction of this type of workers staying at the �rm, F (nz1), rises. The �rm then has a
conventional monopsony problem in which there is a marginal cost of labor that lies above
the labor supply curve and a marginal revenue from an additional unit of labor which, in
this case, is horizontal. The equilibrium cuto� is the one that ensures that marginal cost is
equal to marginal revenue and the equilibrium wage ensures that this cuto� is on the labor
supply curve.
An increase in cL for this worker (which would result if the employee had an s1 equal to
15
zero and �w fell) shifts the labor supply curve to the left because it raises what the worker
can expect to earn outside and thereby makes departures more attractive. This leads to a
reduction in the cuto�, so that more workers of this type leave in equilibrium and, at least
when labor supply is linear, the equilibrium wage increases. An increase in �w lowers cL for
workers with s1 equal to zero while raising it for those with s1 equal to one, so that it moves
the labor supply curve of employees with s1 equal zero to the right while it moves that of
employees with s1 equal to one to the left. It thus increases the relative turnover and wages
of high s1 workers.
Before closing this section, it is worth computing wages and turnover in the case where
F is uniform for n between 0 and 1=h. This means that, when the optimal second period
wage is interior, (2) reduces to
wz2 =
qS + qN + vrz
2(24)
and this applies as long as the second inequality in (3) is satis�ed, or
qS + qN + vrz< 2=h: (25)
Equation (11) then implies that expected second period compensation is
cz2 = h(wz
2 � qN )wz2 +
Z 1=h
wz2�qN
(qN + n)hdn:
Carrying out the integration, this is
cz2 = qN +
h
2(wz
2 � qN )2 +
1
2h(26)
This falls with h when wz2 is equal to qN because an increase in h lowers the mean of the
value of outside o�ers.
Substituting the wage in (24) into (26) (21) becomes
h
hqS + vr
z + h
�Æz2
�2� w
z1
i� h
hwz1 +
h
2
�Æz2
�2+
1
2h� c
zL
i= 0
where Æz = qS � qN + vrz is the di�erence between the expected output of an employee with
s2 = z at a type S and a type N �rm.
16
The optimal �rst period wage is thus
2wz1 = qS + vr
z �1
2h+ c
zL +
h
2
�Æz2
�2: (27)
which is clearly larger for z equal to one.
2 Aggregate \Output"
In this section I show that increases in �w can reduce the sum of the output produced by
�rms and the nonpecuniary utility enjoyed by individuals. I denote the per capita value of
this aggregate output in period t by t. Even when changes in �w raise t, this by no means
implies that such changes lead to Pareto improvements. Indeed, they generally do not since
some workers see their wages rise while others see their wages decline. Still, it is useful to
know the reasons why increasing trust in capitalism by increasing �w can lower, in some
sense, the total value of the pie that is produced.
Given that a fraction � of individuals receive a signal s1 equal to one, per capita output
in period 1 is equal to
1 = (1��)
"F (n01)(qS + vr
0) +
Z �n
n01
(qN + n)dF (n)
#+�
"F (n11)(qS + vr
1) +
Z �n
n11
(qN + n)dF (n)
#:
(28)
Thus, the change in 1 is equal to
d1 = (1� �)[Æ0 � n01]f(n
01)dn
01 + �[Æ1 � n
11]f(n
11)dn
11: (29)
This change depends only on the two cuto�s n01 and n11. An increase in either cuto� leads to
a gain in output that is proportional to the di�erence between the marginal employee's total
output (inclusive of nonpecuniary compensation) in a type S �rm and in a type N �rm. The
existence of a monopsonistic distortion in �rms of type S together with its absence in �rms
of type N implies that this di�erence is positive because workers get paid less than their
marginal product in the former. As a result,
Proposition 6: If F is uniform between 0 and 1=h, d1
d�w< 0 as long as
0 <Æ1
2+h
8(Æ1)2 +
h(1� �)
16
241�
� � �
1� �
!235 [(Æ1)2 � (Æ0)2] <
1
h: (30)
17
If (30) is violated while
0 <Æ0
2+h
8(Æ0)2 +
h�
16
241�
� � �
1� �
!235 �(Æ0)2 �minf
2
h2; (Æ1)2g
�<
1
h; (31)
d1
d�w> 0. If the second inequality in (31) is violated as well, no employee departs �rms of
type S in the �rst period so that �w has no e�ect on output.
Proof: See the Appendix.
Condition (30) e�ectively ensures that some workers with s1 = 1 quit (so n11 < �n) and this
occurs only for certain values of (qS � qN ) and v. Then, employees with both realizations of
s1 alter their quitting behavior as �w changes. Output in period 1 then falls with �w because,
in the uniform case, the di�erence between the marginal employee's output at a �rm of type
S and at a �rm of type N is larger for workers whose s1 equals one. The reason is related
to the result in proposition 2: the di�erence between an employee's wage and his marginal
product is larger for more productive employees so that society has more to lose from having
these employees depart from their type S employer. This then implies that increases in �w
lower aggregate output because the cost that results from the reduction in the attachment of
more productive workers is larger than the gain that results from the increased attachment
of less productive workers.
For other parameter values, particularly those where qS � qN , v or h are relatively large,
(30) can be violated while (31) holds. High values of these parameters imply that �rms of
type S raise wages to raise the probability of keeping their employees. They might then keep
all employees whose s1 equals one while setting wages so that some employees with s1 = 0
depart. Then, only the turnover of employees with s1 = 0 is reduced when �w increases. This
raises output because, at the margin, these employees produce more inside type S �rms.
I now turn my attention to second period per capita output, which I denote by 2.
Proposition 7: In response to changes in n01 and n
11, the change in 2 is given by
d2 = P (s2 = 0js1 = 1)
"Z w12�qN
w02�qN
(vP (r = 1js1 = 1; s2 = 0)� qN � n)dF (n)
#f(n11)dn
11 �
P (s2 = 1js1 = 0)
"Z w12�qN
w02�qN
(vP (r = 1js1 = 0; s2 = 1)� qN � n)dF (n)
#f(n01)dn
01 (32)
18
Proof: Start with the last term. A small reduction in n01 increases by �f(n01)dn
01 the
number of employees with s1 = 0 that leave their original type S employers. If these
employees draw an s2 equal to zero, their output in the second period is the same as if they
had stayed with their original type S employer regardless of the value of their n. If they get
an s2 equal to 1 and their realization of n is either below w02 � qN or above w1
2 � qN , their
total output is again the same because, whether they originally stay or leave, they work
at a type S �rm in period 2 in the former case and at a type N �rm in the latter. Their
period 2 behavior is di�erent only if, after they leave, they get an s2 equal 1 while their
realization of n lies between w02 � qN and w
12 � qN . This leads them to stay at the type S
�rm in period 2 whereas they would have left their original type S �rm if they had stayed in
the �rst period. As a result, such employees see their output increase by qS � qN + vEr� n
where the expectation of r must be computed taking into account that s1 = 0 while s2 = 1.
Similarly, a small reduction in n11 increases by �f(n
11)dn
11 the number of employees with
s1 = 1 that leave their original type S employers. If these individuals draw an s2 equal to 1,
their output is again unchanged regardless of the realization of n. Their behavior changes
only if they draw an s2 equal to zero and an n between w02 � qN and w
12 � qN . This now
leads them to leave their new type S �rm in period 2 whereas they would have stayed with
their original type S �rm if they had stayed in the �rst period. Their output thus falls by
qS � qN + vEr � n where the expectation of r must be computed taking into account that
s1 = 1 while s2 = 0.
Note that the two terms in square brackets in (32) are equal. If these integrals are
positive, 2 falls when �w rises. This, in turn, occurs if individuals whose signals are mixed
(in the sense that they get one signal equal to 1 and one signal equal to 0) and whose
nonpecuniary compensation in type N �rms have values between (w01 � qN ) and (w1
1 � qN)
are more productive on average when they work in type S �rms. That this condition leads
to an output loss can be seen by focusing �rst on individuals with s1 = 1 and s2 = 0. By
leaving their original type S employer, these individuals now work for type N employers in
the second period when n falls in this range and this condition implies that their resulting
19
output is on average lower. This condition implies also that individuals who stay after
receiving an s1 equal to zero lower their social output in period 2 in those states of nature
where they would have gotten an s2 equal to one had they left their original type S employer.
Once again, one can obtain more de�nite results in the uniform case. Then,
Z w12�qN
w02�qN
[vP (r = 1js1 = 1; s2 = 0)�qN�n]dF (n) = h(w12�w
02)
"qS + vP (r = 1js1 = 1; s2 = 0)�
w12 � w
02
2
#
so that 2 falls if qS + vP (r = 1js1 = 1; s2 = 0) � (w12 � w
02)=2, which I denote by �, is
positive. This requires that expected output in a �rm of type S of an employee with a mixed
signal is higher than the arithmetic average between the wages paid to high and low signal
employees in the second period. Moreover,
Proposition 8: If F is uniform, suÆcient conditions for � > 0 include
i) qS � qN + v(1��)�
1+��2�> 0
ii) � < 3=4
iii) � � 3=4 and � � � >4(1��)(��:75)
(4��1)where this latter number is no greater than .026.
There do, however, exist parameters that violate all these conditions and which ensure
that � < 0 so that d2
d�w> 0.
Proof: See the Appendix.
It follows immediately that 2 falls with �w if qS > qN , which is equivalent to assuming
that there exist realizations of n low enough that it is more eÆcient for the employee to stay
at a �rm of type S even if one is sure that his r is equal to zero. Indeed, what is needed to
overturn the conclusion that d2
d�w< 0 is that the productivity of employees in �rms of type
N , qN be relatively high. At the same time, condition (5) which requires that employees
sometimes be more productive in �rms of type S even if their s1 is equal to zero, puts a limit
on qN since it requires that
qN < qS +�(1� �)
1� �
The violations of ii) and iii) in Proposition 8 arise only when � is high and �, while
greater than �, is relatively low. Such combinations of parameters allow qN to be relatively
high without violating (5). It then becomes possible that d2
d�w> 0 because more is produced
20
when employees with mixed signals switch to type N �rms when they draw o�ers whose
value is between w02 and w
12.
That there is a broad range of parameters for which d2
d�w< 0 is not unexpected. Since
�rms pay workers less than their marginal products, it is not too surprising that the average
of w01 and w
11 can be below the expected marginal product of a worker with mixed signals.
Increases in �w then lower 2 in the uniform case. This conclusion can be overturned if
productivity in sector N is suÆciently high, because the monopsony distortion is lower in
this case. Faced with employees whose productivity in sector N is high, �rms of type S
raise wages and thereby reduce the extent to which productivity in sector S exceeds wages.
Still, because productivity is always at least somewhat greater than wages, the range of
parameters such that d2
d�w> 0 is relatively small.
To summarize the results of this section, increases in �w have the potential to lower
output in both periods both because they reduce the most productive attachments in period
1 and because they reduce the extent to which individuals who either get or would have
gotten mixed signals work in type S �rms in the second period. If (30) and any of the three
conditions in Proposition 8 hold, aggregate output actually falls in both periods. This is the
case for a broad range of parameters, including for example, setting qS, qN , v and h equal
1 while � = :7 and � = :9. When v is increased beyond about 1.6, (30) ceases to hold and
1 rises with �w. Another method for violating (30) is to raise h. By raising h to 4, keeping
qS and qN equal to 1, and setting �, � and v to .9, .89 and .462 respectively, the gains in 1
from raising �w exceed the losses in 2.
It is also straightforward to �nd numerical values such that all the conditions of proposi-
tion 8 are violated and 2 rises with �w. However, because this occurs only when Æ0 is close
to zero, h must be small to ensure that the �rst inequality of (31) holds. In my numerical
experiments these low values of h made it impossible to violate (30). I was unable to �nd
parameters for which output falls in both periods.
21
3 Empirical Relevance
The model predicts that populations that di�er only in their beliefs about the extent to
which pay re ects productivity will have di�erent distributions of both turnover and wages.
Unfortunately, data on the extent to which people feel that compensation is tied to pro-
ductivity do not appear to exist, and this limits the extent to which one can ascertain the
importance of these beliefs for labor market outcomes. While these beliefs are not observ-
able, other variables that would seem to be closely related to these beliefs can be measured.
In particular, individuals in di�erent countries di�er both at a point in time and over time
in the programs of their elected representatives and in the typical responses they give to
surveys asking about the desirability of income redistribution. In the absence of improved
data, I treat countries where redistribution is favored (and whose elected representatives
re ect this) as being more skeptical about the extent to which di�erences in compensation
re ect di�erences in productivity. There may well be other reasons why people see inequality
as excessive. They may, for example, regard inequality as a problem because the see poor
individuals as unable to meet what are regarded as basic needs. However, Hochschild (1981,
p 111, 140) reports that both her rich and her poor (U.S.) respondents favored tying income
to individual productivity and di�ered mainly in the extent to which they regarded this as
occurring in practice.
I �rst study the connection between beliefs and inequality and then I turn to the con-
nection between inequality (or beliefs) and turnover. The data on beliefs is more extensive
at a point in time for di�erent countries, and so I start with this. I then discuss the little
bit of data that is available about changing beliefs through time.
Figure 3 focuses on seven countries at a point in time. It shows the relationship be-
tween the attitudes towards inequality from the 1987 International Social Survey as well as
inequality measures from around the same period.10
The vertical axis of this �gure gives the fraction of people who agree with the statement
10Inequality measures change slowly over time so that it should not matter very much that these measure-
ments were not all taken precisely in 1987.
22
that \Di�erences in income in [respondent's country] are too large". These �gures are drawn
from Evans (1993).11 On the horizontal axis, the �gure gives the Gini coeÆcient for household
pre-tax income. These �gures are drawn from Deininger and Squire (1995) who report
inequality measures from a number of sources. To help make these �gures comparable, I
used �gures from the LIS data base ( Atkinson, Rainwater and Smeeding, 1995) for all the
countries for which these �gures were available, namely for all countries except Austria and
Australia.12 The correlation between these �gures is -.72. Such a negative correlation �ts
with the model presented earlier if one regards the opinion that inequality is too large as
closely related to the opinion that individual pay does not re ect individual performance.
There are, of course, alternative interpretations. One possibility is that the perception
of large inequities leads to government policies that reduce actual inequality. In an extreme
version of this interpretation, perceptions a�ect incomes only through their e�ect on policies.
One way these policies could be leading to the negative correlation in �gure 3 is by a�ecting
the measured income distribution itself. The measure of pre tax household income on which
my Gini coeÆcients are based includes some government transfers and cash transfers to poor
people which might be larger in countries where people dislike inequality more. I have thus
also looked at OECD measures of the distribution of labor earnings. The coverage of these
data is not uniform for the seven countries in my sample. For Austria, the U.K. and the U.S.,
the �gures appear to be comparable and the ratio D9/D5 is largest in the U.S., intermediate
in the U.K. and smallest in Austria. The D5/D1 ratio is highest in the U.S. and lower for
the U.K. and Austria, though the latter two numbers are essentially identical. These �gures
are broadly consistent with the Gini coeÆcients for pre-tax income.
Another source of potential endogeneity, is that high-income people might work less
11The only other country included in this international survey was Hungary which I excluded because I
felt that the data on actual inequality was probably not very meaningful during its post-communist trans-
formation.12I picked, for each country the observation that was closest to 1987. The data for Switzerland are from
1982, those for Germany are from 1983, those for Norway, the U.K. and the U.S. are from 1986 and those for
Ireland and the Netherlands are from 1987. The Australian data come from the 1989 Statistical Yearbook
for Australia while those for Austria, whose data are probably the least comparable to the others, comes
from 1987.
23
in countries where perceived inequality is higher because tax rates are more progressive.13
In practice, it is not clear that the countries in my sample in which inequality is seen as
particularly excessive have particularly progressive income taxes. The correlation of the
Gini coeÆcients based on after tax household income (all of which come from the LIS data
base) with the Evans' measure is only -.40.
It may also be possible to rationalize the �ndings in Figure 3 by using the gift exchange
model of Akerlof (1982). In that model, worker e�ort depends on the relationship between
a worker's wage and his \reference wage," which Akerlof supposes to be the wage earned
by similar workers. If the extent to which people see inequality as excessive is related to
workers' reference wages, this dislike could a�ect equilibrium wages. In particular, one might
imagine that societies where inequality is regarded as more excessive are ones where low-
paid workers have higher reference wages (because they see themselves as more similar to
higher paid workers) while workers who are paid more have lower reference wages. This is
not enough to ensure that wage dispersion is lower in such societies. Dispersion falls if �rms
have more to gain from raising wages of low wage workers (and less to gain from raising
wages to high wage workers) in such societies. This might occur if, for example, workers
whose wage is below their reference wage increase their e�ort by more in response to an
increased gift of the �rm than do workers whose wage is already above the reference wage.
It might also occur if increasing the wages of high-paid workers discourages e�ort of low-
paid workers in the same �rm. While this idea still requires development, it suggests that
an e�ort based eÆciency wage model in which norms of gift giving play a role might have
similar implications than those of the turnover-based eÆciency wage model developed here.
Because it corresponds most closely to the belief that pay does not re ect performance,
I focused attention on the survey which asked whether people thought income disparities
were excessive. It is important to note, however, that the answer to this question is strongly
positively correlated with the answer to questions that ask whether people favor additional
13Higher tax rates alone might account for the phenomenon if these have a bigger e�ect on the e�ort of
individuals whose income is high.
24
redistribution from the rich to the poor. Evans (1993) reports a measure which gives the
mean response to questions asking people whether they would like the unemployed to receive
a guaranteed standard of living, whether they support more spending on bene�ts for the
poor, whether they support a guarantee of jobs for everyone and whether they are in favor
of a basic income for everyone. With a higher means score representing less enthusiasm for
redistribution, the correlation between this mean score and the answer plotted in Figure 3
is -.80. Similarly, the mean score against redistribution has a correlation of .78 with the
pre-tax Gini coeÆcient and of .52 with the after-tax coeÆcient.
Time series evidence on attitudes towards income inequality seems even harder to obtain.
This is particularly surprising given that many political analysts have noted a \shift to the
right" in U.S., British and Canadian electoral outcomes. In all three countries, the late 1970's
and early 1980's saw the election of political leaders like Ronald Reagan, Margaret Thatcher
and Brian Mulroney whose rhetoric was substantially more \pro-market" than that of their
predecessors. The only U.S. data that con�rms that this was accompanied by a change in
people's attitude towards redistribution are reported in Kluegel and Smith (1986). They
conducted a national survey in 1980 and found that only 18% of their respondents disagreed
with the statement that the U.S. was spending too much on welfare. As they say, this is
substantially smaller than the 39% of the people who agreed, in a 1969 survey conducted
by Feagin (1975), that the U.S. was spending too little on welfare. This comparison su�ers
from several diÆculties not the least of which is that the public perception of government
policies might have been di�erent in the two instances.
In spite of the paucity of attitudinal data, it may be reasonable to suppose that the change
in electoral outcomes in these three countries is suggestive of an increased con�dence in the
free market system as well as a more benign view of employers.14 It would then be possible
14Other changes in measured attitudes are consistent with this. As Uchitelle (1994) reports citing a study
by Richard Freeman and Joel Rogers, workers no longer view big business as an adversary. They do not
blame their employers for the anxiety that they feel about their income and job security. Instead, workers
regard their employers as victims of global competition which forces employers to cut costs and lay o�
workers.
Thus, interestingly, international competition may have increased income inequality via two mechanisms.
The �rst and more traditional one, is that this competition may have reduced the demand for low skilled
25
to explain the increased income inequality in these three countries over the last 20 years by
this change in attitudes. What makes this explanation attractive is that, as Moss (1995)
shows, inequality has worsened disproportionately in these three countries (together with
Australia). By contrast countries in continental Europe such as France, Italy and Germany
have seen their earnings inequality fall. At the same time, political leaders in these countries
appear not to have shifted nearly as much towards a free market rhetoric.
A more distinctive source of evidence on the importance of the links highlighted in this
paper is the evolution of turnover. Admittedly, the model has only two periods so that taken
literally, its implications concern only whether individuals stay in jobs for one period or not.
However, it seems reasonable to imagine that suitable extensions imply that job durations
are shorter whenever the model implies that people are more likely to separate after one
period. Suppose, in particular, that �w rises. The model implies that more workers with
s1 equal to 1 quit, and this presumably corresponds in practice to relatively rapid turnover
(and thus short jobs) for individuals whose wages are relatively high. I thus take the model
to imply that increases in �w ought to increase job tenure of low wage workers while reducing
that of high wage workers. If other forces are changing everyone's tenure equally, it should
still be the case that the tenure of low wage workers ought to rise relative to that of their
higher paid counterparts.
In practice, tenure is reduced not only by quits (as in my model) but also by employ-
ment terminations that are initiated by employers - most of which take the form of layo�s.
McLaughlin (1991) shows that layo�s are signi�cantly more likely for low education than for
high education workers. This suggests that my model ought to be more consistent with the
evolution of employment durations among individuals with relatively high levels of education.
Naturally, even if actual tenure has changed in the way predicted by my model, this
could be due to other changes in the labor market, including changes in the composition of
the labor force, in the average educational level, in unionization, in managerial practices or
workers. The second is that the existence of this international competition may have convinced workers that
their employers are not being unfair when they cut wages. This, in turn, may have exacerbated inequality
through a mechanism akin to the one presented in this paper.
26
in labor market regulations. The e�ects of these changes on the duration of attachments
between workers and �rms is unknown and I am thus unable to control for them in my
analysis. What is known is that average tenure appears to have fallen in the United States,
particularly among individuals with low levels of educational attainment. The overall size
of this reduction has been small, however. (See Farber (1995), Diebold, Neumark and
Polsky (1996) and Swinnerton and Wial (1996)). In Great Britain, Burgess and Rees (1998)
show that the recent period of vast changes in labor markets has had led to no secular
changes in expected tenure once one controls for demographic, educational and occupational
characteristics of workers.
To study whether the implications for tenure predicted by my model are borne out, I con-
sider data from both the United States and France. Table 1 reports statistics obtained from
the May 1979, January 1991 and February 1996 U.S. Current Population Survey. In particu-
lar, it tabulates means for the answer to the question of how many years the respondent has
worked for his or her current employer. What makes these surveys particularly useful is that
they simultaneously asked for information on earnings. For each of these surveys I �rst di-
vided the data into cells corresponding to gender, to four educational attainment categories
and to four age categories. For each of these 32 categories, I then computed the median
level of hourly earnings, the average job tenure for those whose hourly earnings are below
the median for the cell and the average tenure for those whose hourly earnings are above. I
label those with earnings higher than the median for their cell as high income - recognizing
that this abuses the language somewhat since those with low educational attainment tend
to have incomes that are low relative to the population as a whole even if their income is
above the median for their cell.15 Table 1 presents, for each gender-education-age-income
category, the average tenure in the three survey dates.
Table 2 presents analogous data from France. These data come from the Enqu�ete Em-
15I control for education and other observable characteristics in this way both because these variables are
known to a�ect tenure and because the model's income di�erences are due to signals that only employers
observe. In an appendix available upon request, I also consider an extension which shows the model can
also imply that increases in �w raise inequality across individuals that di�er in observable characteristics like
educational attainment.
27
ploi which was conducted yearly from 1982 to 1998 and includes observations on tenure,
monthly earnings and hours. Francis Kramarz kindly computed median hourly earnings
for 140 gender-education-age categories as well as mean tenure for those earning above and
below this level of earnings. For the purpose of displaying these data, I have taken means
across subcategories so that the age and education categories in the Table are somewhat less
disaggregated.
Consistent with the literature on earnings and seniority (see, for example Abraham and
Farber 1987), workers whose wages are relatively high given their age and education tend
to have spent more time with their current employer. My model implies this as well, as
individuals with s1 = 1 receive a higher wager that reduces their probability of quitting.16
Interestingly, higher levels of education are only weakly associated with higher levels of
tenure (for given age).17
Consistent with the �ndings of Burgess (1998), average tenure levels tend to be consider-
ably longer in France than in the U.S. This is true both for high income individuals (which
is consistent with a lower �w in France) and for low income individuals. This latter fact
is not consistent with the model I presented above if di�erences in �w are the only di�er-
ences across countries. However, this fact might be consistent with an extension where �rms
make inferences about the quality of their employees from the fact that these have left their
previous employer. If high-wage individuals tend to stay with their employers the pool of
job changers is of lower quality on average and this might promote suÆciently low wages
for job changers that low-wage individuals are also more reluctant to change jobs. Whether
an extension along these lines can explain these relative tenure levels is a topic for further
research. I now focus on the changes in tenure over time in the two countries.
16An alternative view is that this correlation results from the automatic escalation of wages with seniority.
See Abraham and Farber (1987) for a discussion and evidence against this alternative view.17For example, U.S. males between 35 and 44 who have attended college for some time have generally
had lower completed spells of employment with their current employer than have males of the same age
who have only complete high school. Still, there is a sense in which, on average, more educated individuals
have slightly longer levels of tenure. In particular, consider the ratio for a given gender-age category of the
average tenure of individuals in the two high education categories to the average tenure of the individuals
in the corresponding two low-education categories. The average ratio of this sort (across gender and age
categories as well as across years) is 1.04.
28
Table 3 provides some simple transformations of the data in Table 1 which permit one
to visualize the changes in tenure both from 1979 to 1991 and from 1991 to 1996.18 The
�rst eight columns shows, for all the gender-age-education-income categories, the ratios of
average tenure in 1991 to the average tenure in 1979 as well as the corresponding ratios for
1991-1996. With rare exceptions it is not the case that average tenure grew for low income
individuals and shrank for high income individuals. A weaker implication of the model is
that tenure for di�erent categories of individuals ought to have grown di�erentially. This is
analyzed in the last two columns. These give, for each category, the di�erence between the
log growth rate of tenure for high wage individuals from 1979 to 1996 and the corresponding
log growth rate for low-wage individuals. The theory implies that this ought to be negative
in the United States.
This turns out to be true when one looks at females as a whole. It is also true on
average for individuals who have stayed in school beyond high school whether they be male
or female. Moreover, high-wage high-education individuals saw their tenure fall relative to
that of low-wage high-education individuals not only overall from 1979 to 1996 but also over
each of the two subperiods. The opposite turns out to be the case for individuals with less
education.19 As I suggested above, layo�s account for a disproportionate number of the
separations of employees with low levels of education so that my model ought to be more
relevant for high-education individuals. Whether the distinction between quits and layo�s
is responsible for these contrary results awaits further research, however.
Interestingly, French tenure for individuals with high educational attainment evolved
rather di�erently. This can be seen in the last two columns of Table 2. These show that, on
average, tenure of high wage individuals (whether male or female) rose relative to tenure of
their low-wage counterparts from 1982 to 1998 when one looks at groups with relatively high
levels of education. The same is true for individuals with lower educational attainment, but
18According to Bernstein and Mishel (1997) inequality in the United States continued to grow in the
second subperiod.19The �rst two columns do show, however, that the tenure of high-income, low-education females grew less
fast than the tenure of low-income low-education females in the period 1979-1991 (with the opposite being
true in the period 1991-1996).
29
this is not distinctive to France since a similar pattern is observed in the United States for
males, and for females from 1991 to 1996.
A broader view of the evolution of French tenure by income can be obtained by considering
regressions which explain the evolution of a variable I call TENURERATIOit. This variable
equals the di�erence between the logarithm of average tenure at t of individuals within group
i whose hourly earnings exceed median hourly earnings for the group and the logarithm of
average tenure of those whose earnings are below the median. I consider, in particular, a
regression of the form
TENURERATIOit = ci + � t+ �it
where ci is a �xed e�ect for the group while � is the coeÆcient on a time trend. A positive
� thus indicates that tenure of high-wage individuals has been growing over time relative to
tenure of low-wage individuals. Table 4 shows that, for various speci�cations, the estimated
value of � is indeed positive. Interestingly, this coeÆcient is not very sensitive to the level
of education of the groups that I include in the regression. In particular, it is positive and
signi�cantly di�erent from zero also for groups with relatively high educational attainment
and this provides a contrast with the U.S. results.
The estimate of � tends to be higher when one drops earlier observations. The highest
t-statistic for � is obtained when one uses only the observations between 1988 and 1998, and
I show the resulting estimate in Table 4. Samples that drop even more early observations
tend to have even higher estimates (though the associated t-statistic is lower because the
number of observations falls). Still, the point estimate using data only from 1982 to 1988 is
positive as well.
The estimates from the subsample with somewhat less formal education are somewhat
larger, particularly for women. One way to compare these results to those using U.S. data is
to multiply the point estimate of � by 17 to compute the expected growth in the log di�erence
in tenure between high and low wage individuals. In the case of women of relatively low
educational attainment this expected growth is .19, which vastly exceeds the corresponding
log di�erence in Table 3, which is only .04. Thus, for these women, it appears that tenure has
30
been lengthening disproportionately for high wage workers in France, as one would predict if
�w rose more in the United States. On the other hand, the point estimates predict a growth
in the log di�erence of tenure for males with relatively low education of only .07 and this is
much smaller than .201, the growth in this di�erence in the U.S. over the period 1979-1996.
The movements in tenure that are most consistent with those predicted by the model
remain those of highly educated individuals. Among this group, high wage workers have been
reducing their tenure relative to low wage workers while the opposite is true in France. An
alternative interpretation for this �nding is that, for exogenous reasons, more opportunities
have become available for high-wage high-education individuals in the United States and that
these opportunities have led to \job-hopping." While this alternative cannot be dismissed,
it should be kept in mind that an increase in �w which leads high-wage workers to be willing
to move may thereby encourage the endogenous creation of these opportunities.
4 Conclusions
This paper presented an attempt at explaining widening income disparities in countries such
as the United States by changes in the extent to which people believe that higher rewards to
more productive individuals. It shows that these changes in beliefs can have a direct e�ect
on the distribution of income even if they do not a�ect the inherent productivity or the e�ort
of any worker.
For these beliefs to matter, employees must �nd it diÆcult to infer the wages they would
ultimately obtain at alternate employers. This suggests that it would be worthwhile to model
the extent to which individuals acquire this knowledge. Early in individuals' careers, there
is a force which can lead individuals to acquire a reasonably large amount of information
about wages at alternate employers. Suppose, as seems plausible, that individuals start out
with in ated opinions about their abilities. If their �rst employer o�ers them a high wage
after the initial screening period, they are likely to conclude that pay is highly correlated
with productivity while they are likely to conclude the opposite if their �rst job o�ers a
31
low remuneration.20 Both of these inferences lead individuals to be quite willing to try
alternative employers. This, in turn, leads them to acquire hard information both about
their opportunities and about the extent to which wages o�ered at di�erent employers are
correlated with each other (which helps them calibrate employer accuracy). It would be
worthwhile to understand these dynamics better. It seems reasonable to suppose that the
cost of these moves makes this process end before individuals feel they know the relevant
parameters. Moreover, as Tables 1 and 2 indicate, many individuals end up staying with
single employers for a long time. Employers are likely to base their wage o�ers to these
individuals on what these individuals expect to get if they do leave and these expectations
may well rely increasingly on general information, including opinions about the accuracy of
employers.
While the model assumes that beliefs are homogeneous - so as to focus on common
changes in beliefs - more work deserves to be done under the assumption that communication
is suÆciently diÆcult that beliefs remain dispersed. This dispersion is certainly consistent
with the survey evidence. One might thus be able to use survey evidence to study whether
individuals whose beliefs di�er also act di�erently. Is it the case, for example, that high
wage workers who believe that pay is not strongly correlated with productivity have stayed
at their current jobs for a relatively long time?
In closing I ought to emphasize again that one advantage of the model is that it allows
one to evaluate the extent to which changes in beliefs are responsible for widening income
disparities in di�erent countries by studying whether changes in tenure are consistent with
these changed beliefs. I have looked at measures of average tenure for workers which fall
within certain sex-age-education and income cells in the both the United States and in
France and have found some support for the model. Clearly, much more deserves to be done
along these lines. It would be better, for example, to compare the separation behavior of
individuals taking into account both the history of their wages and the likely future wages
at their employer. This might be done using data that identi�es employees and employers
20This contrast is consistent with the observations of Hochschild (1981) mentioned earlier.
32
simultaneously as in Abowd, Kramarz and Margolis (1999). They show that employers di�er
in the extent to which they make their wages depend on seniority. Since workers ought to
take this into account when they decide to quit, it would be useful to know whether quitting
behavior has changed in the ways suggested by my model when this deferral of compensation
is taken into account.
33
5 References
Abowd, John, Francis Kramarz and David Margolis, \High Wage Workers and High WageFirms," Econometrica, 67, March 1999, 251-333.
Acemoglu, Daron and J�orn-Ste�en Pischke, \Why do Firms Train? Theory and Evi-dence," MIT Department of Economics Working paper 96-7, February 1996.
Akerlof, George A., \Labor Contracts as Partial Gift Exchange," Quarterly Journal of
Economics, 97, November 1982, 54-69.
Atkinson, Anthony B., Lee Rainwater, and Timothy Smeeding, Income Distribution inEuropean Countries, Working Paper Series, No 121, Luxembourg Income Study,1995
Bernstein, Jared and Lawrence Mishel, \Has Inequality Stopped Growing?," Monthly
Labor Review, 120, December 1997, 3-16.
Bolton, Patrick and Christopher Harris, \Strategic Experimentation,"Econometrica, March1999, 349-74.
Burgess, Simon, \The Reallocation of Labour: An International Comparison using JobTenure Data", mimeo 1998.
Burgess, Simon and Hedley Rees, \A Disaggregated Analysis of the Evolution of JobTenure in Britain, 1975-1993," British Journal of Industrial Relations, 36, Decem-ber 1998, pp. 629-55.
Deininger, Klaus and L. Squire, \Measuring Income Inequality: A new Data-Base,"World Bank mimeo, 1995.
Diebold, Francis X., David Neumark and Daniel Polsky, \Comment on Kenneth Swin-nerton and Howard Wial \Is Job Stability Declining in the U.S. Economy?" "Industrial and Labor Relations Review, 49, January 1996, pp. 348-352.
Evans, Geo�rey, \Class con ict and Inequality," in Jowell, Robert, Lindsay Brooks andLizanne Dowds, coeditors, International Social Attitudes, the 10th BSA Report,
Brook�eld VT: Dartmouth Publishing Co., 1993.
Farber, Henry, \Are Lifetime Jobs Disappearing? Job Duration in the United States:1973-1993", National Bureau of Economic Research Working Paper 5014, February1995.
34
Feagin, J.R. Subordinating the Poor, Englewood cli�s, N.J.: Prentice-Hall, 1975.
Hochschild, Jennifer, What's Fair? American Beliefs about Distributive Justice, Cam-bridge, MA: Harvard University Press, 1981.
Kluegel, James R. and Eliot R. Smith, Beliefs about Inequality, New York: De Gruyter,1986.
Lundberg, Shelly and Richard Startz, \Private Discrimination and Social Intervention inCompetitive Labor Markets," American Economic Review, 73, June 1983, 340-47.
McLaughlin, Kenneth J., \A Theory of Quits and Layo�s with EÆcient Turnover," Jour-nal of Political Economy, 99, February 1991, 1-29.
McClosky, Herbert and John Zaller, The American Ethos, Cambridge MA: HarvardUniversity Press, 1984.
Moss, David A., Unemployment in France: \Priority Number One", Case 9-795-064,Boston, MA: Harvard Business School Publishing, 1995.
Piketty, Thomas, \Social Mobility and Redistributive Politics," Quarterly Journal of
Economics, 110 (3), August 1995, pp 551-84.
Swinnerton K. and H. Wial, \Is Job Stability Declining in the U.S. Economy? Reply toDiebold, Neumark and Polsky" Industrial and Labor Relations Review, 49, January1996, 352-355.
Uchitelle, Louis, \Insecurity forever," New York Times, 11/20/94.
35
Appendix
Proof of Proposition 6:I �rst show that (31) implies that Æ0 < 2h so that w0
2 is given by (24). This can be seenby noting that, because � < 1 and � � �, the second inequality of (31) implies that
Æ0
2+h
4(Æ0)2 <
3
2h
The second inequality in (30) implies both the second inequality of (31) and Æ1< 2=h so
that w12 is also given by (24). Using (17) and (26), we thus have
c1L � c
12 =
h(1� �)
8
241�
� � �
1� �
!235 [(Æ0)2 � (Æ1)2]:
This means that, using (19), (24) and (27), n11 is given by
n11 =
Æ1 + c
12 � c
1L
2+h
8(Æ1)2 =
Æ1
2+ (A.1)
h
8(Æ1)2 +
h(1� �)
16
241�
� � �
1� �
!235 [(Æ0)2 � (Æ1)2]:
This means that (30) is equivalent to the condition that 0 < n11 < 1=h. When this
condition holds, changes in �w a�ect n11. By contrast, n11 simply remains equal to 1=h (or 0)
in response to local changes in �w if the condition is strictly violated.The analogous argument when s1 equals zero (so that we use (16) to compute (c0L � c
02)
implies that, when (30) holds so that second period wages are given by (24)
n01 =
Æ0 + c
02 � c
0L
2+h
8(Æ0)2 =
Æ0
2+ (A.2)
h
8(Æ0)2 +
h�
16
241�
� � �
1� �
!235 [(Æ1)2 � (Æ0)2]
which, given (30) implies that n10 is smaller than 1=h as well. Thus, assuming the expressionin (A.2) is strictly positive, this cuto� also changes with �w. Given that both cuto�s changeaccording to (23), replacing f by h and f
0 by zero, (29) becomes
d1
d�w=
�h(�w � �)
1� �(c12 � c
02)n[Æ0 � n
01]� [Æ1 � n
11]o: (A.3)
Using (A.1) and (A.2), the term in braces is
h
16[(Æ1)2 � (Æ0)2]
241 + 1
2
0@1�
� � �
1� �
!21A35� v
2(r1 � r
0):
36
When r1 equals r0, this is zero, The derivative of this expression with respect to r1 is
v
8<:h4(Æ1)
241 + 1
2
0@1�
� � �
1� �
!21A35� 1
2
9=; � v
"3h
8(Æ1)�
1
2
#
< v
"6h
8h�
1
2
#< 0
where the �rst inequality follows from the fact that � is no smaller than � while the secondfollows from the fact that Æ1 is smaller than 2=h. Thus d1
d�wis negative when (30) holds.
Note that, if the expression in (A.2) is negative, all employees with s1 equal to zero leavetheir �rst employer so that n01 does not vary with �w. Thus, a fortiori, increases in �w lower1.
Now consider the case where (30) fails while (31) holds. Since it is no longer certain thatÆ1< 2=h (though this inequality is true when Æ
1 is replaced by Æ0), (16) and (26) now implythat
c0L � c
02 =
h�
8
241�
� � �
1� �
!235 [minf 4
h2; (Æ1)2g � (Æ0)2]:
Equations (19) and (27) thus imply that
n01 =
Æ0
2+h
8(Æ0)2 +
h�
16
241�
� � �
1� �
!235 [minf 4
h2; (Æ1)2g � (Æ1)2]:
Thus, (31) ensures that 0 < n01 < 1=h so that changes in �w still a�ect n01. Therefore,
d1
d�w=
�h(�w � �)
1� �(c12 � c
02)[Æ
0 � n01]:
Using (A.2), the term in brackets is
Æ0
2
1� Æ
0h
4
!+c0L � c
02
2:
This is greater than zero because the last term is positive while Æ0 is less than 2=h. Thisestablishes that 1 rises with �w when (30) fails while (31) holds. A small variant on theabove calculations demonstrates that neither n01 nor n
11 varies with �w when both the second
inequality in (30) and the second inequality in (31) fail.
Proof of Proposition 8: From Bayes rule
P (r = 1js2 = 0; s1 = 1) =P (r = 1; s2 = 0js1 = 1)
P (s2 = 0js1 = 1)=
P (s2 = 0jr = 1; s1 = 1)P (r = 1js1 = 1)
P (s2 = 0js1 = 1)
=�(1� �)(1� �)
(1� �)2 � (� � �)2=
�(1� �)
1 + � � 2�: (A.4)
37
Thus, given (5) and (24)
� = qS + vP (r = 1js1 = 1; s2 = 0)�w
12 � w
02
2�
qS � qN
2+ v
"�(1� �)
1 + � � 2���
4��(1� �)
4(1� �)
#:
A strict inequality obtains if the second boundary condition in (3) is violated because thiswould mean that smaller wages are suÆcient to keep workers with probability one. Thus
qS � qN + v(1��)�
1��
2v+
�(1� �)
1 + � � 2���
4�3(1� �)�
4(1� �)> 0) � = 0: (A.5)
Multiplying through by (1 + � � 2�)(1� �) and manipulating, this condition becomes
2f(1��)((qS�qN )(1+��2�)=v+(1��)�)+(1��)�(���)g+(���)[3�6�+�(4��1)] > 0:
Because � must exceed � and the zeros of the equation 4�2 � 7� + 3 = 0 occur at 3/4and 1, the term in square brackets is positive if � is less than 3/4. Since (5) ensures thatthe term in braces (which is proportional to the �rst term in (A.5)) is nonnegative, this issuÆcient to imply that � > 0. The term in square brackets can also be written as
(� � �)(4�� 1)� 4(1� �)(3=4� �):
It is thus also positive when � > 3=4 as long as iii) is satis�ed. Moreover, for � > 3=4, theratio 4(1� �)(3=4� �)=(4�� 1) reaches a maximum of .0253 when � equals .862. The lastinequality above can also be written as
2f(1� �)(1 + � � 2�)
((qS � qN) +
v(1� �)�
1 + � � 2�
)+ (� � �)[3� 4�+ �(2�� 1)] > 0:
Given that the zeros of the equation 2�2� 5�+ 3 = 0 occur at 1 and at 1.5 whereas � mustbe less than one, the term in brackets must be nonnegative. Therefore, it is suÆcient for� > 0 that the term in braces be positive.
When all three conditions are violated while h is also small enough that (25) holds for
both types of employees, d1
d�w< 0.
38
Appendix IIObservable Characteristics
In the body of the paper I have treated the signals s1 and s2 as observed privately
by �rms and their employees. Thus the increase in �rst period wage inequality I derived
must be thought of as an increase in the inequality within groups with identical observable
characteristics. Such increases have been observed in the United States (Levy and Murnane
1992) and their existence among college graduates poses something of a challenge to the
view that increased inequality is simply the result of declines in the demand for low skilled
labor which have been brought about by increases in computer use and international trade.
It turns out, however, that one can extend the model so that it explains increased inequality
across individuals who di�er in the value of a publicly observable signal U . I demonstrate
this by �rst extending the model and then considering a special case in which, while U is
positively correlated with s1, s2 and r, it is irrelevant to the agents in the model because their
own signals (s1 and s2) are much better indicators of r. To an outside observer, however,
the di�erence in average income between individuals with high and low values of U increases
whenever the income of individuals whose s1 equals one rises relative to the income of those
whose s1 equals zero.
To show this, I start by ignoring the publicly observable signal U and describe a relatively
general model of the connection between s1, s2 and r. As in Figure 2, each worker has a
type i which can take eight distinct values which I denote by A, B, C, D, a, b, c and d.
Productivity r equals one if i equals a capital letter (and otherwise equals zero). The signal
s1 equals one if i equals A, a, B or b while s2 equals one if i equals A, a, C or c. In Figure
2, the outcomes with s1 = 1, s2 = 1 and r = 1 are depicted as overlapping circles.
Letting �i denote the probability that a worker's outcome takes the value indicated by i,
the conditional and unconditional probabilities that play a key role are
P (r = 1) = �A + �B + �C + �D
P (s1 = 1) = �A + �B + �a + �b P (s2 = 1) = �A + �C + �a + �c
39
P (r = 1js1 = 1) =�A + �B
�A + �B + �a + �bP (r = 1js2 = 1) =
�A + �C
�A + �C + �a + �c
P (r = 1js1 = 0) =�C + �D
�C + �D + �c + �dP (r = 1js2 = 0) =
�B + �D
�B + �D + �b + �d
P (s2 = 1js1 = 1) =�A + �a
�A + �B + �a + �bP (s2 = 0js1 = 0) =
�D + �d
�C + �c + �D + �d:
By setting the �rst three unconditional probabilities to � and the following two condi-
tional probabilities to �, one can solve for the seven independent values of �i after imposing
the two independence assumptions in (13). I do not consider that special case here, however.
Rather, I let the �i's perceived by employers remain constant while I vary the beliefs of
workers. Suppose workers believe that �b and �c increase while �d declines correspondingly.
This means that workers perceive s1 and s2 to be weaker signals so that their assessments
of P (r = 1js1 = 1 and P (r = 1js2 = 1) decline. The same is true of P (s2 = 1js1 = 1) and
P (s2 = 0js1 = 0) so that the two signals are seen as less correlated with one another. This
has the same e�ect as the decline in �w considered above: it makes workers with s1 = 1
remain with higher probability with their �rst type S employer while those with s1 = 0
become more keen to leave.
Now imagine that, for each value of i, the publicly observable signal U is equal to one
with probability �i. For �rms to ignore the value of this signal, it must be the case that
the six conditional probabilities above are independent of whether U is equal to zero or one.
The �rst of these requires that P (r = 1js1 = 1) = P (r = 1js1 = 1; U = 1) or,
�A + �B
�A + �B + �a + �b=
�A�A + �B�B
�A�A + �B�B + �a�a + �b�b
which can also be written as
(�a + �b)(�A�A + �B�B)� (�A + �B)(�a�a + �b�b) = 0:
The requirements that P (r = 1js2 = 1), P (r = 1js1 = 0), P (r = 1js2 = 0), P (s2 = 1js1 =
1) and P (s2 = 0js1 = 0) be independent of U can similarly be written, respectively, as
(�a + �c)(�A�A + �C�C)� (�A + �C)(�a�a + �c�c) = 0
40
(�c + �d)(�C�C + �D�D)� (�C + �D)(�c�c + �d�d) = 0
(�b + �d)(�B�B + �D�D)� (�B + �D)(�b�b + �d�d) = 0
(�B + �b)(�A�A + �a�a)� (�A + �a)(�B�B + �b�b) = 0
(�a + �b)(�A�A + �B�B)� (�A + �B)(�a�a + �b�b) = 0
This can be written compactly as
M� = 0 (A.6)
where M is a 6�8 matrix of coeÆcients that depend on the �i's and � is the vector given by
[�A �B �C �D �a �b �c �d]0.
Now consider the position of an outside observer. This observer would see wages of
individuals with U = 1 rising relative to wages of individuals with U = 0 when wages of
employees with s1 = 1 rise relative to those of employees with s1 = 0 if a high value of U is
more prevalent among individuals with s1 = 1. This occurs if
P (s1 = 1jU = 1) =�A�A + �B�B + �a�a + �b�bP
i �i�i> �A + �B + �a + �b = P (s1 = 1)
or
(�C+�D+�c+�d)(�A�A+�B�B+�a�a+�b�b)�(�A+�B+�a+�b)(�C�C+�D�D+�c�c+�d�d) > 0
(A.7)
In general, (A.7) is inconsistent with (A.6), though there is a broad set of �i's such that
they both hold simultaneously. In particular,
Proposition 9 If M is of full row rank, (A.7) holds as an equality. By contrast, if M is of
rank 5, one can �nd �i's such that both (A.6) and (A.7) hold.
Sketch of Proof Inspection of the above equations shows that one can write each of the
columns of M as linear combinations of the other columns with coeÆcients equal to +1 and
-1. When arrayed in the same order, the coeÆcients of the �'s in (A.7) can be written as the
same linear combination of the other coeÆcients of (A.7). This means that, ifM has full row
rank, there exists a linear combination of the rows ofM which has the same eight coeÆcients
41
as (A.7). Equation (A.6) implies that the inner product of this linear combination and the
vector [A B : : : c d]0 equals zero.
Even when M has full row rank, one can set two �'s to arbitrary values and �nd values
for the remaining �'s that satisfy (A.6). In general, these need not be between zero and
one. However, if the two �s that are chosen arbitrarily are set equal to one another, the only
solution to (A.6) is to set all �'s equal to this common value. By continuity, setting the two
arbitrary �'s both between 0 and 1 and close to one another ensures that the remaining �'s
are between zero and one as well.
If M does not have full row rank, there generally is no linear combination of rows of M
with coeÆcients equal to those of (A.6). This means that one can �nd �'s such that (A.6)
holds with the left hand side being equal to any positive number. Moreover, if the left hand
side is set equal to zero, and one sets two �'s equal to the same arbitrary value, the trivial
solution to the combination of (A.6) and (A.7), is to set all other �'s to this value as well.
This means that, by continuity, one can �nd �'s between zero and one that ensure that the
left hand side of (A.7) is equal to a small positive number.
The proposition shows that, in general, knowing that the signal is uninformative to �rms
about the wages that they ought to pay as well as being uninformative to workers about the
wages they can expect to receive implies that it is uncorrelated with wages. The case where
this is not true is when the �'s are such that the lack of informativeness of U for some agents
follows automatically from its lack of informativeness for others. The requirement that the
U 's be uninformative to the agents in the model then imposes fewer restriction on the �'s so
that it becomes possible to make U informative about s1 (and r).
This is easiest to illustrate in the special case where
�A�c = �a�C and �B�d = �b�D: (A.8)
The �rst of these conditions requires that the odds of r = 1 given that s2 = 1 be independent
of s1. Similarly, the second requires that the odds of r = 1 given that s2 = 0 be independent
of s1.
42
Now suppose that �A, �B, �a and �b are all set equal to the same constant �� > 0 while
all other �i's are set equal to zero. Thus, U is greater than zero only if s1 = 1 so that wages
are clearly correlated with U . On the other hand, it is easy to check that all the conditions
in (A.6) are satis�ed. First, U does not convey information about r given s1 because I have
chosen the values of �i so that they depend only on s1 and not on r given s1. It also does
not convey information about r given s2 because (A.8) ensures that the odds of r given that
both s1 and s2 equal one are the same as if only s2 equals one. Since U is one only when s1
is one and conveys no information about r given s1, it also conveys no further information
about r given s2. Lastly, the reason U does not convey information about s2 given s1 is that,
if s1 = 0, U must be zero as well while, if s1 = 1, (A.8) implies that the odds of s2 = 1 are
the same whether U equals one or not.
This is by no means the only possible outcome when (A.8) holds. It is also possible to
choose �'s so that the di�erence between the probability that U = 1 given that s1 = 1 and
the probability that U = 1 given that s1 = 0 is much smaller. For example, I computed
numerically the solutions for � that result from setting �A = �a = :15, �B = �C = �c =
(�b=2) = :08 and �D = (�d=2) = :1 and letting the left hand side of (A.7) be equal to .01.
After arbitrarily setting �c=.25 and �d = :26, the remaining �'s ensure that P (U = 1js1 = 1)
is about .27 while P (U = 1) is only slightly above .25.
It is worth stressing that I am not suggesting that the assumptions in (A.8) are generally
realistic.21 Rather, it is likely that the proper assumptions regarding the �i's depend on
whether the Us represent gender, education, or other observable characteristics such as
race. Hopefully, this appendix provides a basis for further research on the e�ect of these
characteristics on wages.
Levy, Frank and Richard J. Murnane, \U.S. Earnings Levels and Earnings Inequality:A Review of Recent Trends and Proposed Explanations," Journal of Economic
Literature, 30, September 1992, 1333-81.
21In fact, (A.8) is inconsistent with the assumptions I made earlier about the probabilities of s1, s2 and
r. Consider again the case where P (s1 = 1) = P (s2 = 1) = P (r = 1) = �, s1 and s2 are independent of
each other given r and P (r = 1js1 = 1) = P (r = 1js2 = 1) = �. Then, �B=�D > �b=�d. The former equals
�=(1� �) while the latter equals �=(1� � +(���)(2��)
(1��)2)
43
Type S firm:Output = qS+vrWage = w1
z ifsignal s1=z
Type N firm:Output=wage=qNDisutility n
Period 1 Period 2
Type S firm:Output = qS+vrWage = w2
z ifsignal s1=z
Type N firm:Output=wage=qNDisutility n
Type S firm:Output = qS+vrWage = w2
z ifsignal s2=z
Figure 1The sequence of potential employers
Supply
Supply’
MCMC’
W*’
W*
n
w
MR
Figure 2Wage Determination
Figure 3Actual and perceived inequality
United StatesAustralia
Switzerland
United Kingdom
Netherlands
Germany
Austria
50
55
60
65
70
75
80
85
90
31 33 35 37 39 41
Pre-tax Gini Coefficient
Per
cept
ion
of e
xces
sive
ineq
ualit
y
Figure 4Observable Characteristics
r=1: A U D U Fs=1: A U B U D U Ee=1: A U B U C
G
F
D E
A B
C
MALES FEMALES
Low income High income Low income High income79 91 96 79 91 96 79 91 96 79 91 96
High School not completedAge between 25 and 34 3.53 2.24 3.08 4.87 4.13 4.68 2.02 2.92 1.58 3.59 4.07 3.61Age between 35 and 44 6.28 6.06 4.54 8.81 9.52 8.90 3.78 2.76 2.50 6.24 5.79 5.86Age between 45 and 54 10.21 9.06 8.79 15.03 15.47 15.01 6.14 6.58 6.56 9.73 13.06 11.07Age between 55 and 64 11.75 9.29 8.10 18.11 17.38 19.51 7.96 8.15 10.57 13.25 14.83 14.53
High School completedAge between 25 and 34 3.82 3.56 3.79 5.43 5.88 5.72 2.53 3.08 2.85 4.85 4.94 5.71Age between 35 and 44 8.48 6.70 6.88 10.23 10.95 10.55 4.10 4.40 4.66 7.26 8.18 8.30Age between 45 and 54 13.31 10.63 10.27 17.76 16.09 16.93 5.90 6.85 7.10 11.40 10.57 12.99Age between 55 and 64 14.25 9.83 10.11 20.99 18.81 18.21 8.87 9.23 8.00 15.50 13.54 14.76
Attended CollegeAge between 25 and 34 3.47 3.56 2.98 5.14 5.49 4.40 2.65 3.13 2.78 4.26 4.76 4.34Age between 35 and 44 6.87 7.54 6.70 10.12 9.24 9.93 3.03 4.31 5.44 6.60 8.76 8.08Age between 45 and 54 11.76 10.35 10.42 16.49 14.91 14.47 6.63 5.67 7.14 10.65 10.64 11.53Age between 55 and 64 12.85 10.51 10.04 20.34 19.45 15.38 7.89 9.28 10.43 13.96 16.21 15.21
Post-College educationAge between 25 and 34 3.17 3.34 3.80 4.48 4.77 3.78 2.37 3.10 3.39 4.27 3.72 3.47Age between 35 and 44 7.56 7.26 6.31 9.48 9.20 8.60 5.06 6.50 6.88 8.01 8.33 7.83Age between 45 and 54 11.82 12.56 11.00 15.03 14.02 14.06 8.75 8.83 9.67 12.56 11.74 14.85Age between 55 and 64 14.49 16.27 11.69 18.53 19.02 14.82 13.44 11.25 12.61 16.94 16.32 16.92
Table 1The Levels of Tenure in the United States
Low High Low High1982 1996 1982 1996 1982 1996 1982 1996 Income Income Income Income
No diploma and CEP5.52 4.25 6.43 5.98 5.46 3.97 6.27 6.13 0.77 0.93 0.73 0.988.80 9.63 11.83 13.82 7.38 6.76 9.88 12.40 1.09 1.17 0.92 1.2511.77 14.23 17.70 20.24 10.73 11.96 14.34 23.60 1.21 1.14 1.11 1.6514.41 15.09 19.75 21.90 12.99 14.45 16.22 18.58 1.05 1.11 1.11 1.14
CAP and BEPC6.59 4.89 8.11 6.73 6.72 4.16 8.97 6.74 0.74 0.83 0.62 0.7512.05 10.56 14.17 15.10 8.97 8.59 13.89 15.59 0.88 1.07 0.96 1.1215.96 16.04 21.54 22.43 12.43 12.92 20.65 22.98 1.01 1.04 1.04 1.1117.15 19.02 23.13 22.01 13.10 14.69 22.68 22.73 1.11 0.95 1.12 1.00
Baccalaureat5.96 4.14 7.18 5.92 5.92 4.37 7.83 6.61 0.69 0.82 0.74 0.8511.41 10.04 13.01 14.23 10.97 9.72 14.10 15.09 0.88 1.09 0.89 1.0716.34 18.03 20.68 21.90 12.50 15.20 21.35 22.70 1.10 1.06 1.22 1.0617.64 17.52 25.64 20.26 16.79 16.33 27.12 25.47 0.99 0.79 0.97 0.94
University degree3.75 2.62 4.25 3.51 3.86 3.13 5.10 3.91 0.70 0.83 0.81 0.7710.45 8.72 10.48 9.52 10.41 8.99 11.02 10.92 0.83 0.91 0.86 0.9914.80 16.78 16.67 18.23 14.12 16.80 21.29 18.97 1.13 1.09 1.19 0.8922.89 21.10 21.68 22.49 15.38 20.42 27.79 27.51 0.92 1.04 1.33 0.99
0.94 0.99 0.98 1.04
0.98 1.03 0.95 1.13
0.91 0.95 1.00 0.94
Males FemalesLow Income High Income Low Income High Income
Low Education Average
Age between 26 and 35Age between 36 and 45Age between 46 and 55Age between 56 and 65
Age between 56 and 65
Age between 26 and 35
Females
Unweighted Average
Age between 26 and 35Age between 36 and 45Age between 46 and 55
Age between 36 and 45Age between 46 and 55Age between 56 and 65
Levels of Tenure Tenure 1996 / Tenure 1982
High Education Average
Table 2The Evolution of Tenure in France
Males
Age between 56 and 65
Age between 26 and 35Age between 36 and 45Age between 46 and 55
Low Income
High Income
Low Income
High Income
Low Income
High Income
Low Income
High Income
1979 -Males
1996 Females
High School not completedAge between 25 and 34 0.64 0.85 1.45 1.13 1.37 1.14 0.54 0.89 0.095 0.246Age between 35 and 44 0.96 1.08 0.73 0.93 0.75 0.93 0.91 1.01 0.335 0.350Age between 45 and 54 0.89 1.03 1.07 1.34 0.97 0.97 1.00 0.85 0.148 0.062Age between 55 and 64 0.79 0.96 1.02 1.12 0.87 1.12 1.30 0.98 0.446 -0.192
High School completedAge between 25 and 34 0.93 1.08 1.22 1.02 1.06 0.97 0.92 1.16 0.059 0.046Age between 35 and 44 0.79 1.07 1.07 1.13 1.03 0.96 1.06 1.01 0.239 0.006Age between 45 and 54 0.80 0.91 1.16 0.93 0.97 1.05 1.04 1.23 0.212 -0.055Age between 55 and 64 0.69 0.90 1.04 0.87 1.03 0.97 0.87 1.09 0.201 0.053
Attended CollegeAge between 25 and 34 1.03 1.07 1.18 1.12 0.84 0.80 0.89 0.91 -0.004 -0.032Age between 35 and 44 1.10 0.91 1.42 1.33 0.89 1.07 1.26 0.92 0.006 -0.382Age between 45 and 54 0.88 0.90 0.85 1.00 1.01 0.97 1.26 1.08 -0.009 0.005Age between 55 and 64 0.82 0.96 1.18 1.16 0.96 0.79 1.12 0.94 -0.033 -0.194
Post-College educationAge between 25 and 34 1.05 1.07 1.31 0.87 1.14 0.79 1.09 0.93 -0.351 -0.565Age between 35 and 44 0.96 0.97 1.28 1.04 0.87 0.94 1.06 0.94 0.084 -0.329Age between 45 and 54 1.06 0.93 1.01 0.93 0.88 1.00 1.10 1.26 0.005 0.067Age between 55 and 64 1.12 1.03 0.84 0.96 0.72 0.78 1.12 1.04 -0.009 0.063
Unweighted Average 0.91 0.98 1.11 1.06 0.96 0.95 1.03 1.02 0.074 -0.073
Low Education Average 0.81 0.98 1.10 1.06 1.01 1.01 0.95 1.03 0.201 0.039
High Education Average 1.00 0.98 1.13 1.05 0.91 0.89 1.11 1.00 -0.043 -0.179
growth rates by incomeLog difference in tenure
Table 3The Evolution of Tenure in the United States
Tenure 1991 / Tenure 1979 Tenure 1996 / Tenure 1991Males Females Males Females
Table 4 Regression Results
Speci�cation � Std. Error. Number of obs. R2 (within)
All Observations� .0060 .00097 2193 0.019
High-educationy .0042 .0014 963 0.011
High-education, 1988-1998 .0130 .0025 623 0.045
High education, 1982-1987 .0045 .0063 340 0.002
Low-education, males .0042 .0016 613 0.012
Low-education, females .0109 .0022 617 0.040
Notes: � Excludes two groups which included either two non consecutive observations or one observation. The remaining groups contain at
least seven observations and all of their observations are consecutive.
y Includes individuals who have either completed their baccaleureat or continued going to school after high-school.